What is Artificial Intelligence?

Science fiction literature of the 1940s greatly helped inspire scientists and engineers of the time on the topic of artificial intelligence. A crucial role was played by American writer Isaac Asimov, who in his 1942 short story Runaround devised the Three Laws of Robotics : 1) A robot may not injure a human being or,

In the 1950s, British mathematician Alan Turing was the first intellectual to seriously question artificial intelligence. "Can machines think?" At the time, there were only four computers in the world (Manchester Mark I, Cambridge EDSAC, ENIAC, BINAC). The debate on machine intelligence had just begun. The answer to this question turned out to be too

American philosopher John Searle was the most important antagonist of Alan Turing regarding philosophical thought on AI. According to Searle, machines may appear intelligent, but they have neither consciousness nor real understanding. To demonstrate this, he devised the "Chinese Room" thought experiment: The Chinese Room : Imagine a closed room. Inside is a person who

During the First (1760–1840), Second (1870–1914), and Third Industrial Revolutions (1970–today), new technologies were invented that changed how tasks were performed, boosting productivity, innovation, and enthusiasm for new possibilities. We’re talking about the steam engine, railways, and coal as a fossil fuel in the first; electricity, the light bulb, and the internal combustion engine in

Amara's Law states that people tend to overestimate the effect of a technology in the short term and underestimate it in the long term. The Hype Cycle, on the other hand, describes how, when a new technology is introduced to the market, it follows five phases: The five phases of the Hype Cycle: 1) Innovation

Machine Learning is only one of the possible approaches to making a machine intelligent; today it is the most suitable to obtain most of the desired results (outputs). However, it is not always the most functional approach to use, since Machine Learning requires, in addition to specific skills, large amounts of data and computational power.

Machine Learning marked a real break from earlier alternative methods, as it allows machines to handle data never seen before. Unlike in the past, humans do not necessarily have to program the machine with specific rules (e.g., rule-based systems, etc.); after training see Module 4 , the machine is able to make predictions or generate

In this approach, humans organize the data delivered to the machine, making it easier for the machine to link inputs (incoming data) with outputs (expected results), since it starts from correctly labeled data. There are two types of supervised learning: 1) Classification Learning → The machine must generate a categorical output. Example: Training an algorithm

In this approach, humans do not organize the data delivered to the machine; rather, the machine must analyze and organize the data on its own based on similarities (clustering). Therefore, unlike supervised learning, the machine learns only from inputs (what in supervised learning is called the Phase 1 - training set ), without having outputs

With this approach, the machine learns by experimenting: 1) It tries actions: The machine attempts to perform an action in the surrounding environment (for example, in a video game, it tries moving to the right); 2) It receives feedback (positive or negative) from the environment: It receives a "reward" or a "penalty" from the environment,

Deep Learning consists in using artificial neural networks (ANNs), or neural nets , within Machine Learning to allow the machine to learn even better from data. ANNs are software inspired by the human brain, albeit in a simplified way. In the human brain, biological neurons receive signals from dendrites (inputs). When the sum of these

The artificial neural network (ANN), as anticipated, enables the machine to learn in an even deeper way (Deep Learning) compared to “classic” Machine Learning. A neural network is composed of three main parts: Input layer : receives raw incoming data (e.g., images, text, numbers, etc.) and vectorizes them; Hidden layers : progressively process and transform

In this deep dive we will analyze the MLP (Multilayer Perceptron) in detail—a simple Deep Learning architecture that helps understand the logic behind all other artificial neural networks, which we will briefly review at the end. Let’s assume a machine must learn from data using a neural network and that, among the input data (potentially

Deep Learning enables results that classic Machine Learning methods (Supervised Learning, Unsupervised Learning, and Reinforcement Learning) struggle to achieve, especially when the data are very large, complex, or unstructured (images, audio, text). In particular, Deep Learning allows the machine to analyze and generate outputs for which it was not explicitly programmed (Generative AI or GenAI

Thanks to Generative AI, we moved from using computers to classify items, guess sequences, and apply rules, to using them to generate brand-new, high-quality content such as text, images, code, audio, and video—all without being explicitly programmed to do so. Generative Artificial Intelligence (Generative AI or GenAI ) is an application paradigm (we’ll clarify this

The term "AI Model" refers to the result of machine learning from data through Machine Learning (whether supervised, unsupervised, or reinforcement learning). An AI Model becomes complex and capable of producing Generative AI only through Deep Learning, where the many hidden layers and numerous nodes enable a better understanding of data compared to "classic" Machine

An AI Agent is software that connects a human with an AI Model. AI Agents allow humans to use AI Models for specific tasks (Chat, Image Creation, Software Development, specific vertical features for business, etc.). They do this by managing: Features, Parameters, and Prompts. Features: all the activities an agent must perform before, during, and

In this deep dive we’ll get practical: the operational steps needed to build an AI Model: 1) Choosing the IDE An IDE is software that collects in one program all the tools useful for writing, modifying, testing, and managing an application’s code. The most famous are Visual Studio Code , IntelliJ , NetBeans , Eclipse

A prompt is text that conveys an instruction or a question to the AI Model in order to obtain an output. The more the prompt is optimized for the AI Model, the better the output generated by the Model will be. Example: if I wanted to know which cities have the highest number of residents

Fine-tuning consists of taking a pre-trained Model on a very large general dataset and "refining" it for a specific purpose by training it further on a smaller, specialized dataset. A Model fine-tuned for a specific task almost always beats a generic Model or a Model trained from scratch on a small dataset. It leverages the

Fine-tuning can also be achieved through Reinforcement Learning from Human Feedback ( RLHF ), where people evaluate the Model’s answers by assigning a score (or rank) to the best responses. The Model learns to predict which response would be rated better by humans ( Reward Model ) through a reinforcement learning algorithm (e.g., PPO –

Ethics is the branch of philosophy that studies moral principles and values, distinguishing between right and wrong behavior, good and bad actions, etc., with the goal of guiding moral decisions. The ethical dimension is extremely sensitive in AI and is often broken down into: • Ethical AI : AI systems that respect ethical principles and

Algorithmic bias occurs when an algorithm produces skewed results due to biases present in the training data. Machine biases can stem from human prejudices present in data and also from confirmation bias , i.e., when a human gives more weight to information that confirms their beliefs, thereby influencing the machine. Humans can also be influenced

As we saw earlier, Explainable AI refers to AI in which the machine’s decision can be understood and explained. One of the main problems with Generative AI is, in fact, that it’s not possible to understand the logic behind a decision. The stochastic nature of these new Generative AI Models does not help, since the

Every new technology brings opportunities but also risks. These risks necessarily require regulation to prevent direct or indirect harm to people or property. Regulation exists to ensure safety and trust in the adoption of the technology. The first comprehensive regulation approved in the field of artificial intelligence was the Artificial Intelligence Act , passed by

The limits of an AI Model depend not only on data quality, but also on the machine’s computational power, which in turn is influenced by the complexity of the underlying problem. Polynomial time , or P , is the set of problems that a computer can solve in polynomial time, i.e., with time that grows

Nondeterministic Polynomial time , or NP , is the set of problems for which, given a solution, a computer can verify in polynomial time whether it is correct. However, finding the optimal solution from the problem statement may require exponential time. Example: imagine a computer must find the winning combination of a safe with 4

Beyond computational limits, there are also data limits. We previously analyzed qualitative limits with algorithmic bias, but that is not the only data limit—there is also a quantitative one. In fact, the amount of data required to obtain a valid output depends on the number of variables we want to analyze. More variables require more

After reading all the deep dives on this site, you’ll be able to assess whether an AI Model is too expensive for your professional needs, what AI’s limits are, what an AI Agent is and how to leverage it, what a RAG is, what fine-tuning is, the difference between open-source and closed software, privacy matters,

As we have seen, AI lets you easily solve all P -type problems—at least those that do not require generating substantial outputs or analyzing very large amounts of inputs. First, you should identify all the time-consuming, low value-added activities that take up part of your day—regardless of your role. Once those tasks are automated, you